work-on-task: --non-interactive mode for headless callers

CHANGELOG.md

@@ -11,6 +11,7 @@ Please choose versions by [Semantic Versioning](http://semver.org/).
 ## Unreleased
 
 - refactor(plugin): rename orchestration-mode flag `--tool` → `--non-interactive` and internal label `MODE=tool` → `MODE=non_interactive` across create-task, create-goal, complete-task, complete-goal, defer-task, defer-goal commands and task-creator, goal-creator, goal-manager-agent agents; the command/agent instructions still accept `--tool` as a deprecated alias for one release. These flags are Claude Code slash-command arguments interpreted by the agent instructions — not vault-cli binary flags.
+- feat(plugin): `/vault-cli:work-on-task` accepts `--non-interactive` — still orients the task (assistant sets status, tracks the daily note, discovers guides) but skips the create-task prompt (Phase 4) and the plan-task/execute-task sharpening chain (Phase 5), so headless callers (e.g. `vault-cli work-on`'s `claude --print` bootstrap) orient and stop instead of hanging on `AskUserQuestion`
 
 ## v0.96.0
 

commands/work-on-task.md

@@ -1,6 +1,6 @@
 ---
 description: Find task details, transition Jira, set status, track on daily note, discover guides, then auto-sharpen and gate the planning → execution transition
-argument-hint: "<jira-id-or-text>"
+argument-hint: "<jira-id-or-text> [--non-interactive]"
 allowed-tools: [Task, AskUserQuestion, Skill, Bash(vault-cli *)]
 ---
 
@@ -16,14 +16,17 @@ Find task details and relevant operational guides before starting work. Delegate
 
 ## Process
 
-1. **Validate input**
-   - If no argument: `❌ Pass a task identifier or description.` and STOP
+1. **Parse input**
+   - Parse `$ARGUMENTS`: if it contains `--non-interactive` → set `MODE=non_interactive` and strip that flag token from the arguments; otherwise `MODE=interactive`. Parsing is self-contained here — it does not depend on any other command. Use the stripped arguments as the task identifier everywhere below — NEVER pass the flag token into the assistant prompt or task search.
+   - If no argument remains after stripping: `❌ Pass a task identifier or description.` and STOP
+
+   **`MODE=non_interactive`** exists for headless callers — e.g. `vault-cli work-on`'s Claude bootstrap runs `claude --print`, which cannot answer `AskUserQuestion`, so an interactive gate would block until the session-start timeout. In this mode the command NEVER calls `AskUserQuestion` and NEVER runs the interactive sharpening chain (Phase 5): it orients (via the assistant) and stops. Interactive sharpening resumes later when the session is opened in a real terminal.
 
 2. **Invoke work-on-task-assistant**
    ```
    Task tool with:
      subagent_type: 'vault-cli:work-on-task-assistant'
-     prompt: 'Find details and guides for: {arguments}'
+     prompt: 'Find details and guides for: {stripped arguments}'
    ```
 
 3. **Auto-sharpen + auto-gate (Phase 5).** If the assistant's report ends with `Ready to work on this task.` (the `found` case), continue to Phase 5 below. If the report contains the `not_found:` marker, skip Phase 5 and run Phase 4 (Handle not_found) instead.
@@ -45,6 +48,8 @@ Output ends with `Ready to work on this task.`
 
 The agent (dispatched in `## Process` step 2) emits a structured `not_found` verdict from its own Phase 1 (`Find task`) when the requested task cannot be found in any source. This phase parses that verdict and asks the user before any file is created.
 
+**Non-interactive gate (checked first):** If `MODE=non_interactive`, do NOT call `AskUserQuestion` and do NOT create anything. Print the `not_found:` report — the `Searched:` block from the verdict, then `❌ Task not found: "<input>"` — followed by `ℹ️ Non-interactive mode: no task created. Re-run in a terminal to create one.` and STOP. Skip steps 1–7 below (they are the interactive create-gate).
+
 1. **Parse the agent's report** for the `not_found:` marker AND capture the verdict body into variables. The agent's `<output_format>` defines two separate fenced markdown blocks — one for the `found` case (ends with `Ready to work on this task.`) and one for the `not_found:` case (literal `not_found:` header on its own line). Look for the `not_found:` block specifically; if the report ends with `Ready to work on this task.` and contains no `not_found:` block, Phase 4 is a no-op and you are done. When the `not_found:` block IS present, match on the `not_found:` token, then extract:
    - `SEARCHED_BLOCK` — the bullet list under the `Searched:` line (verbatim, line-by-line, until the next blank line or `Suggested task name:` line)
    - `SUGGESTED_NAME` — the value after `Suggested task name:` (verbatim, trimmed)
@@ -78,6 +83,8 @@ Goal: by the time work-on-task returns, the resolved task is in `phase: planning
 
 Runs only after Phase 2 returned a `found` task — never on `not_found` (Phase 4 handles that branch).
 
+**Non-interactive gate (checked first):** If `MODE=non_interactive`, SKIP Phase 5 entirely — do NOT invoke `plan-task` or `execute-task`. Both own `AskUserQuestion` flows and would hang a headless caller. The assistant's orient (status, daily note, guides) already ran in `## Process` step 2 and IS the complete non-interactive result. Print `✅ Oriented (non-interactive). Run /vault-cli:plan-task in a terminal to sharpen and gate execution.` and STOP.
+
 1. **Resolve the task name from the assistant's report.** The assistant prints `📋 Task: <name>` near the top of its `found` block; that line is the canonical identifier. Capture it verbatim.
 
 2. **Invoke `Skill: vault-cli:plan-task` with the captured name as a quoted argument.** plan-task owns its own entry contract, gate logic, and `AskUserQuestion` flow — do not intercept. Wait for it to return.